Merge pull request #34788 from BerriAI/litellm_fix_s3_vectors_search

fix(vector_stores): s3 vectors search router bypass + rag query config drop + ui error swallow

litellm/llms/azure_ai/vector_stores/transformation.py

@@ -19,6 +19,7 @@ from litellm.types.vector_stores import (
 
 if TYPE_CHECKING:
     from litellm.litellm_core_utils.litellm_logging import Logging as _LiteLLMLoggingObj
+    from litellm.router import Router
 
     LiteLLMLoggingObj = _LiteLLMLoggingObj
 else:
@@ -115,6 +116,7 @@ class AzureAIVectorStoreConfig(BaseVectorStoreConfig, BaseAzureLLM):
         litellm_logging_obj: LiteLLMLoggingObj,
         litellm_params: dict,
         extra_body: dict[str, Any] | None = None,
+        router: "Router | None" = None,
     ) -> tuple[str, dict[str, Any]]:
         """
         Transform search request for Azure AI Search API

litellm/llms/base_llm/vector_store/transformation.py

@@ -17,6 +17,7 @@ from litellm.types.vector_stores import (
 
 if TYPE_CHECKING:
     from litellm.litellm_core_utils.litellm_logging import Logging as _LiteLLMLoggingObj
+    from litellm.router import Router
 
     from ..chat.transformation import BaseLLMException as _BaseLLMException
 
@@ -57,6 +58,7 @@ class BaseVectorStoreConfig:
         litellm_logging_obj: LiteLLMLoggingObj,
         litellm_params: dict,
         extra_body: dict[str, Any] | None = None,
+        router: "Router | None" = None,
     ) -> tuple[str, dict]:
         pass
 
@@ -69,6 +71,7 @@ class BaseVectorStoreConfig:
         litellm_logging_obj: LiteLLMLoggingObj,
         litellm_params: dict,
         extra_body: dict[str, Any] | None = None,
+        router: "Router | None" = None,
     ) -> tuple[str, dict]:
         """
         Optional async version of transform_search_vector_store_request.
@@ -84,6 +87,7 @@ class BaseVectorStoreConfig:
             litellm_logging_obj=litellm_logging_obj,
             litellm_params=litellm_params,
             extra_body=extra_body,
+            router=router,
         )
 
     @abstractmethod
@@ -197,6 +201,7 @@ class BaseDirectVectorStoreConfig(BaseVectorStoreConfig):
         litellm_logging_obj: LiteLLMLoggingObj,
         litellm_params: Mapping[str, object],
         extra_body: Mapping[str, object] | None = None,
+        router: "Router | None" = None,
     ) -> NoReturn:
         raise NotImplementedError("Direct vector store providers execute the search themselves; no HTTP request shape")
 

litellm/llms/bedrock/vector_stores/transformation.py

@@ -27,6 +27,7 @@ from litellm.types.vector_stores import (
 
 if TYPE_CHECKING:
     from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
+    from litellm.router import Router
 else:
     LiteLLMLoggingObj = Any
 
@@ -196,6 +197,7 @@ class BedrockVectorStoreConfig(BaseVectorStoreConfig, BaseAWSLLM):
         litellm_logging_obj: LiteLLMLoggingObj,
         litellm_params: dict,
         extra_body: dict[str, Any] | None = None,
+        router: "Router | None" = None,
     ) -> tuple[str, dict]:
         if isinstance(query, list):
             query = " ".join(query)

litellm/llms/custom_httpx/llm_http_handler.py

@@ -178,6 +178,7 @@ if TYPE_CHECKING:
         AnthropicMessagesStreamingResponse,
     )
     from litellm.llms.base_llm.passthrough.transformation import BasePassthroughConfig
+    from litellm.router import Router
     from litellm.types.llms.openai_evals import (
         CancelEvalResponse,
         CancelRunResponse,
@@ -2923,7 +2924,7 @@ class BaseLLMHTTPHandler:
         final_response: Final = await self._call_agentic_completion_hooks(
             response=initial_response,
             model=model,
-            messages=(input if isinstance(input, list) else [{"role": "user", "content": input}]),
+            messages=(input if isinstance(input, list) else [{"role": "user", "content": input}]),  # pyright: ignore[reportArgumentType]  # pre-existing mismatch surfaced by the Router import; the hook accepts response input items at runtime
             anthropic_messages_provider_config=responses_api_provider_config,
             anthropic_messages_optional_request_params=response_api_optional_request_params,
             logging_obj=logging_obj,
@@ -5415,7 +5416,7 @@ class BaseLLMHTTPHandler:
         try:
             response: ResponsesAPIResponse | BaseResponsesAPIStreamingIterator = await litellm.aresponses(
                 model=patch.model or model,
-                input=patch.messages,
+                input=patch.messages,  # pyright: ignore[reportArgumentType]  # pre-existing mismatch surfaced by the Router import; patch messages are valid response input at runtime
                 **optional_params,
                 **kwargs_for_followup,
             )
@@ -9688,6 +9689,7 @@ class BaseLLMHTTPHandler:
         timeout: float | httpx.Timeout | None = None,
         client: HTTPHandler | AsyncHTTPHandler | None = None,
         _is_async: bool = False,
+        router: "Router | None" = None,
     ) -> VectorStoreSearchResponse:
         if isinstance(vector_store_provider_config, BaseDirectVectorStoreConfig):
             self._pre_call_direct_vector_store_search(
@@ -9738,6 +9740,7 @@ class BaseLLMHTTPHandler:
                 litellm_logging_obj=logging_obj,
                 litellm_params=dict(litellm_params),
                 extra_body=extra_body,
+                router=router,
             )
         else:
             (
@@ -9751,6 +9754,7 @@ class BaseLLMHTTPHandler:
                 litellm_logging_obj=logging_obj,
                 litellm_params=dict(litellm_params),
                 extra_body=extra_body,
+                router=router,
             )
         all_optional_params: Final[dict[str, object]] = dict(litellm_params)
         all_optional_params.update(vector_store_search_optional_params or {})
@@ -9802,6 +9806,7 @@ class BaseLLMHTTPHandler:
         timeout: float | httpx.Timeout | None = None,
         client: HTTPHandler | AsyncHTTPHandler | None = None,
         _is_async: bool = False,
+        router: "Router | None" = None,
     ) -> VectorStoreSearchResponse | Coroutine[object, object, VectorStoreSearchResponse]:
         if _is_async:
             return self.async_vector_store_search_handler(
@@ -9816,6 +9821,7 @@ class BaseLLMHTTPHandler:
                 extra_body=extra_body,
                 timeout=timeout,
                 client=client,
+                router=router,
             )
 
         if isinstance(vector_store_provider_config, BaseDirectVectorStoreConfig):
@@ -9862,6 +9868,7 @@ class BaseLLMHTTPHandler:
             litellm_logging_obj=logging_obj,
             litellm_params=dict(litellm_params),
             extra_body=extra_body,
+            router=router,
         )
 
         all_optional_params: Final[dict[str, object]] = dict(litellm_params)

litellm/llms/gemini/vector_stores/transformation.py

@@ -33,6 +33,7 @@ from litellm.types.vector_stores import (
 
 if TYPE_CHECKING:
     from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
+    from litellm.router import Router
 else:
     LiteLLMLoggingObj = Any
 
@@ -168,6 +169,7 @@ class GeminiVectorStoreConfig(BaseVectorStoreConfig):
         litellm_logging_obj: LiteLLMLoggingObj,
         litellm_params: dict,
         extra_body: Mapping[str, object] | None = None,
+        router: "Router | None" = None,
     ) -> tuple[str, dict]:
         """
         Transform search request to Gemini's generateContent format.

litellm/llms/milvus/vector_stores/transformation.py

@@ -19,6 +19,7 @@ from litellm.types.vector_stores import (
 
 if TYPE_CHECKING:
     from litellm.litellm_core_utils.litellm_logging import Logging as _LiteLLMLoggingObj
+    from litellm.router import Router
 
     LiteLLMLoggingObj = _LiteLLMLoggingObj
 else:
@@ -123,6 +124,7 @@ class MilvusVectorStoreConfig(BaseVectorStoreConfig):
         litellm_logging_obj: LiteLLMLoggingObj,
         litellm_params: dict,
         extra_body: dict[str, Any] | None = None,
+        router: "Router | None" = None,
     ) -> tuple[str, dict[str, Any]]:
         """
         Transform search request for Azure AI Search API

litellm/llms/openai/vector_stores/transformation.py

@@ -21,6 +21,7 @@ from litellm.utils import add_openai_metadata
 
 if TYPE_CHECKING:
     from litellm.litellm_core_utils.litellm_logging import Logging as _LiteLLMLoggingObj
+    from litellm.router import Router
 
     LiteLLMLoggingObj = _LiteLLMLoggingObj
 else:
@@ -99,6 +100,7 @@ class OpenAIVectorStoreConfig(BaseVectorStoreConfig):
         litellm_logging_obj: LiteLLMLoggingObj,
         litellm_params: dict,
         extra_body: dict[str, Any] | None = None,
+        router: "Router | None" = None,
     ) -> tuple[str, dict]:
         encoded_vector_store_id: Final = encode_url_path_segment(vector_store_id, field_name="vector_store_id")
         url: Final = f"{api_base}/{encoded_vector_store_id}/search"

litellm/llms/pg_vector/vector_stores/transformation.py

@@ -8,6 +8,7 @@ from litellm.types.vector_stores import VectorStoreSearchOptionalRequestParams
 
 if TYPE_CHECKING:
     from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
+    from litellm.router import Router
 else:
     LiteLLMLoggingObj = Any
 
@@ -80,6 +81,7 @@ class PGVectorStoreConfig(OpenAIVectorStoreConfig):
         litellm_logging_obj: LiteLLMLoggingObj,
         litellm_params: dict,
         extra_body: dict[str, Any] | None = None,
+        router: "Router | None" = None,
     ) -> tuple[str, dict]:
         encoded_vector_store_id: Final = encode_url_path_segment(vector_store_id, field_name="vector_store_id")
         url: Final = f"{api_base}/{encoded_vector_store_id}/search"

litellm/llms/ragflow/vector_stores/transformation.py

@@ -17,6 +17,7 @@ from litellm.types.vector_stores import (
 
 if TYPE_CHECKING:
     from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
+    from litellm.router import Router
 else:
     LiteLLMLoggingObj = Any
 
@@ -92,6 +93,7 @@ class RAGFlowVectorStoreConfig(BaseVectorStoreConfig):
         litellm_logging_obj: LiteLLMLoggingObj,
         litellm_params: dict,
         extra_body: dict[str, Any] | None = None,
+        router: "Router | None" = None,
     ) -> tuple[str, dict]:
         """RAGFlow vector stores are management-only, search is not supported."""
         raise NotImplementedError("RAGFlow vector stores support dataset management only, not search/retrieval")

litellm/llms/s3_vectors/vector_stores/transformation.py

@@ -1,8 +1,8 @@
-import re
 from typing import TYPE_CHECKING, Any, Final
 
 import httpx
 
+from litellm.caching._embedding_router import resolve_embedding_router
 from litellm.llms.base_llm.vector_store.transformation import BaseVectorStoreConfig
 from litellm.llms.bedrock.base_aws_llm import BaseAWSLLM
 from litellm.types.router import GenericLiteLLMParams
@@ -18,6 +18,7 @@ from litellm.types.vector_stores import (
 
 if TYPE_CHECKING:
     from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
+    from litellm.router import Router
 else:
     LiteLLMLoggingObj = Any
 
@@ -58,13 +59,20 @@ class S3VectorsVectorStoreConfig(BaseVectorStoreConfig, BaseAWSLLM):
         return headers
 
     def get_complete_url(self, api_base: str | None, litellm_params: dict) -> str:
-        aws_region_name: Final = litellm_params.get("aws_region_name")
-        if not aws_region_name:
-            raise ValueError("aws_region_name is required for S3 Vectors")
-        if not re.match(r"^[a-z][a-z0-9-]*$", aws_region_name):
-            raise ValueError("Invalid aws_region_name format")
+        # Resolve region the same way the ingestion path does:
+        # dynamic param -> AWS_REGION_NAME -> AWS_REGION -> default (us-west-2)
+        aws_region_name: Final = self.get_aws_region_name_for_non_llm_api_calls(litellm_params.get("aws_region_name"))
         return f"https://s3vectors.{aws_region_name}.api.aws"
 
+    def _resolve_query_embedding_router(self, embedding_model: str, router: "Router | None") -> "Router | None":
+        """Return the router iff it serves ``embedding_model`` as a deployment."""
+        if router is None:
+            return None
+        model_list: Final = [
+            dict(m) for m in (router.get_model_list() or ())
+        ]  # mutable-ok: resolve_embedding_router requires list[dict]
+        return resolve_embedding_router(embedding_model=embedding_model, llm_router=router, llm_model_list=model_list)
+
     def transform_search_vector_store_request(
         self,
         vector_store_id: str,
@@ -74,6 +82,7 @@ class S3VectorsVectorStoreConfig(BaseVectorStoreConfig, BaseAWSLLM):
         litellm_logging_obj: LiteLLMLoggingObj,
         litellm_params: dict,
         extra_body: dict[str, Any] | None = None,
+        router: "Router | None" = None,
     ) -> tuple[str, dict]:
         """Sync version - generates embedding synchronously."""
         # For S3 Vectors, vector_store_id should be in format: bucket_name:index_name
@@ -99,10 +108,16 @@ class S3VectorsVectorStoreConfig(BaseVectorStoreConfig, BaseAWSLLM):
 
         # Generate embedding for the query
         embedding_model: Final = litellm_params.get("embedding_model", "text-embedding-3-small")
+        embedding_router: Final = self._resolve_query_embedding_router(embedding_model=embedding_model, router=router)
 
         import litellm as litellm_module
 
-        embedding_response: Final = litellm_module.embedding(model=embedding_model, input=[query])
+        embedding_input: Final = [query]  # mutable-ok: the embedding API takes list input
+        embedding_response: Final = (
+            embedding_router.embedding(model=embedding_model, input=embedding_input)
+            if embedding_router is not None
+            else litellm_module.embedding(model=embedding_model, input=embedding_input)
+        )
         query_embedding: Final = embedding_response.data[0]["embedding"]
 
         url: Final = f"{api_base}/QueryVectors"
@@ -128,6 +143,7 @@ class S3VectorsVectorStoreConfig(BaseVectorStoreConfig, BaseAWSLLM):
         litellm_logging_obj: LiteLLMLoggingObj,
         litellm_params: dict,
         extra_body: dict[str, Any] | None = None,
+        router: "Router | None" = None,
     ) -> tuple[str, dict]:
         """Async version - generates embedding asynchronously."""
         # For S3 Vectors, vector_store_id should be in format: bucket_name:index_name
@@ -153,10 +169,16 @@ class S3VectorsVectorStoreConfig(BaseVectorStoreConfig, BaseAWSLLM):
 
         # Generate embedding for the query asynchronously
         embedding_model: Final = litellm_params.get("embedding_model", "text-embedding-3-small")
+        embedding_router: Final = self._resolve_query_embedding_router(embedding_model=embedding_model, router=router)
 
         import litellm as litellm_module
 
-        embedding_response: Final = await litellm_module.aembedding(model=embedding_model, input=[query])
+        embedding_input: Final = [query]  # mutable-ok: the embedding API takes list input
+        embedding_response: Final = (
+            await embedding_router.aembedding(model=embedding_model, input=embedding_input)
+            if embedding_router is not None
+            else await litellm_module.aembedding(model=embedding_model, input=embedding_input)
+        )
         query_embedding: Final = embedding_response.data[0]["embedding"]
 
         url: Final = f"{api_base}/QueryVectors"

litellm/llms/vertex_ai/vector_stores/rag_api/transformation.py

@@ -21,6 +21,7 @@ from litellm.types.vector_stores import (
 
 if TYPE_CHECKING:
     from litellm.litellm_core_utils.litellm_logging import Logging as _LiteLLMLoggingObj
+    from litellm.router import Router
 
     LiteLLMLoggingObj = _LiteLLMLoggingObj
 else:
@@ -161,6 +162,7 @@ class VertexVectorStoreConfig(BaseVectorStoreConfig, VertexBase):
         litellm_logging_obj: LiteLLMLoggingObj,
         litellm_params: dict,
         extra_body: Mapping[str, object] | None = None,
+        router: "Router | None" = None,
     ) -> tuple[str, dict[str, object]]:
         """
         Transform search request for Vertex AI RAG API

litellm/llms/vertex_ai/vector_stores/search_api/transformation.py

@@ -25,6 +25,7 @@ from litellm.types.vector_stores import (
 
 if TYPE_CHECKING:
     from litellm.litellm_core_utils.litellm_logging import Logging as _LiteLLMLoggingObj
+    from litellm.router import Router
 
     LiteLLMLoggingObj = _LiteLLMLoggingObj
 else:
@@ -245,6 +246,7 @@ class VertexSearchAPIVectorStoreConfig(BaseVectorStoreConfig, VertexBase):
         litellm_logging_obj: LiteLLMLoggingObj,
         litellm_params: dict,
         extra_body: Mapping[str, object] | None = None,
+        router: "Router | None" = None,
     ) -> tuple[str, dict[str, object]]:
         """
         Transform a search request for the Vertex AI Search (Discovery Engine) API.

litellm/proxy/rag_endpoints/endpoints.py

@@ -9,6 +9,7 @@ Provides:
 import base64
 import json
 from collections.abc import Mapping
+from types import MappingProxyType
 from typing import TYPE_CHECKING, Any, Final
 
 import orjson
@@ -19,6 +20,9 @@ from starlette.datastructures import UploadFile
 import litellm
 from litellm._logging import verbose_proxy_logger
 from litellm.constants import DEFAULT_MAX_RECURSE_DEPTH
+from litellm.integrations.vector_store_integrations.vector_store_pre_call_hook import (
+    LiteLLM_ManagedVectorStore,
+)
 from litellm.litellm_core_utils.streaming_handler import CustomStreamWrapper
 from litellm.proxy._types import *
 from litellm.proxy.auth.auth_utils import is_request_body_safe
@@ -36,6 +40,10 @@ from litellm.proxy.rag_endpoints.upload_security import (
     RejectedUpload,
     validate_upload,
 )
+from litellm.proxy.vector_store_endpoints.endpoints import (
+    build_request_data_from_managed_vector_store,
+    reject_caller_embedding_selection_params,
+)
 from litellm.proxy.vector_store_endpoints.utils import (
     assert_user_can_access_vector_store_id,
 )
@@ -120,12 +128,21 @@ def _collect_vector_store_ids_from_payload(payload: object) -> set[str]:
 async def _authorize_nested_vector_store_ids(
     payload: object,
     user_api_key_dict: UserAPIKeyAuth,
-) -> None:
-    for vector_store_id in sorted(_collect_vector_store_ids_from_payload(payload)):
-        await assert_user_can_access_vector_store_id(
-            vector_store_id=vector_store_id,
-            user_api_key_dict=user_api_key_dict,
-        )
+) -> Mapping[str, LiteLLM_ManagedVectorStore]:
+    """Authorize every nested vector store id and return the managed stores it resolved."""
+    return MappingProxyType(
+        {
+            vector_store_id: store
+            for vector_store_id in sorted(_collect_vector_store_ids_from_payload(payload))
+            if (
+                store := await assert_user_can_access_vector_store_id(
+                    vector_store_id=vector_store_id,
+                    user_api_key_dict=user_api_key_dict,
+                )
+            )
+            is not None
+        }
+    )
 
 
 def _build_file_metadata_entry(
@@ -700,11 +717,27 @@ async def rag_query(
                 status_code=400,
                 detail={"error": "retrieval_config must contain 'vector_store_id'"},
             )
-        await _authorize_nested_vector_store_ids(
+        reject_caller_embedding_selection_params(payload=retrieval_config, source="retrieval_config")
+        resolved_stores: Final = await _authorize_nested_vector_store_ids(
             payload=retrieval_config,
             user_api_key_dict=user_api_key_dict,
         )
 
+        # Merge litellm-managed vector store params (provider, region, embedding
+        # model, credentials, ...) from the registry: the same source the direct
+        # /vector_stores/{id}/search endpoint uses. Store-managed keys win on
+        # conflict so callers cannot override the store's provider or credentials.
+        managed_store: Final = resolved_stores.get(retrieval_config["vector_store_id"])
+        store_data: Final = (
+            await build_request_data_from_managed_vector_store(managed_store)
+            if managed_store is not None
+            else MappingProxyType({})
+        )
+        merged_retrieval_config: Final = {
+            **retrieval_config,
+            **store_data,
+        }  # mutable-ok: litellm.aquery requires a plain dict payload
+
         # Add litellm data
         request_data: dict[str, object] = {}
         request_data = await add_litellm_data_to_request(
@@ -716,13 +749,18 @@ async def rag_query(
             proxy_config=proxy_config,
         )
 
-        verbose_proxy_logger.debug("RAG Query - model: %s, retrieval_config: %s", model, retrieval_config)
+        verbose_proxy_logger.debug(
+            "RAG Query - model: %s, vector_store_id: %s, custom_llm_provider: %s",
+            model,
+            retrieval_config["vector_store_id"],
+            merged_retrieval_config.get("custom_llm_provider"),
+        )
 
         # Call query
         response: Final = await litellm.aquery(
             model=model,
             messages=messages,
-            retrieval_config=retrieval_config,
+            retrieval_config=merged_retrieval_config,
             rerank=rerank,
             stream=stream,
             router=llm_router,

litellm/proxy/vector_store_endpoints/endpoints.py

@@ -1,3 +1,5 @@
+from collections.abc import Mapping
+from types import MappingProxyType
 from typing import (
     Annotated,
     Any,  # noqa: TID251  # jsonify_object in proxy/utils.py is annotated with a bare dict
@@ -27,11 +29,69 @@ from litellm.types.vector_stores import IndexCreateRequest, IndexListResponse
 from litellm.vector_stores.vector_store_registry import VectorStoreIndexRegistry
 
 router: Final = APIRouter()
+
+BLOCKED_QUERY_EMBEDDING_SELECTION_PARAMS: Final = frozenset(
+    {
+        "embedding_model",
+        "litellm_embedding_model",
+        "litellm_embedding_config",
+        "litellm_credential_name",
+    }
+)
+
+
+def reject_caller_embedding_selection_params(payload: Mapping[str, object], source: str) -> None:
+    blocked: Final = sorted(BLOCKED_QUERY_EMBEDDING_SELECTION_PARAMS & payload.keys())
+    if blocked:
+        raise HTTPException(
+            status_code=400,
+            detail={
+                "error": f"'{blocked[0]}' cannot be set in {source}. "
+                "Embedding configuration comes from the vector store's server-side registration."
+            },
+        )
+
+
 ########################################################
 # OpenAI Compatible Endpoints
 ########################################################
 
 
+async def build_request_data_from_managed_vector_store(
+    vector_store: LiteLLM_ManagedVectorStore,
+) -> Mapping[str, object]:
+    """
+    Build request params (provider, credential ref, litellm_params) from an
+    already-resolved managed vector store.
+
+    ``litellm_embedding_config`` is resolved here, at request-handling time,
+    instead of at row-creation time: the resolved api_key/api_base/api_version
+    lives only in the returned per-request mapping and is never persisted back
+    to the registry cache. Legacy rows that already carry a resolved
+    (cleartext) config skip the lookup and pass through unchanged.
+    """
+    top_level: Final = MappingProxyType(
+        {
+            key: vector_store.get(key)
+            for key in ("custom_llm_provider", "litellm_credential_name")
+            if key in vector_store
+        }
+    )
+    litellm_params: Final = vector_store.get("litellm_params") or MappingProxyType({})
+    embedding_model: Final = litellm_params.get("litellm_embedding_model")
+    if not embedding_model or litellm_params.get("litellm_embedding_config"):
+        return MappingProxyType({**top_level, **litellm_params})
+
+    from litellm.proxy.proxy_server import prisma_client
+
+    resolved_config: Final = await _resolve_embedding_config(
+        embedding_model=embedding_model, prisma_client=prisma_client
+    )
+    if not resolved_config:
+        return MappingProxyType({**top_level, **litellm_params})
+    return MappingProxyType({**top_level, **litellm_params, "litellm_embedding_config": resolved_config})
+
+
 async def _update_request_data_with_litellm_managed_vector_store_registry(
     data: dict,
     vector_store_id: str,
@@ -51,47 +111,14 @@ async def _update_request_data_with_litellm_managed_vector_store_registry(
     vector_store_to_run: Final[LiteLLM_ManagedVectorStore | None] = await get_litellm_managed_vector_store(
         vector_store_id=vector_store_id
     )
-    if vector_store_to_run is not None:
-        if user_api_key_dict is not None:
-            await assert_user_can_access_vector_store(
-                vector_store=vector_store_to_run,
-                user_api_key_dict=user_api_key_dict,
-            )
-
-        if "custom_llm_provider" in vector_store_to_run:
-            data["custom_llm_provider"] = vector_store_to_run.get("custom_llm_provider")
-
-        if "litellm_credential_name" in vector_store_to_run:
-            data["litellm_credential_name"] = vector_store_to_run.get("litellm_credential_name")
-
-        if "litellm_params" in vector_store_to_run:
-            litellm_params = vector_store_to_run.get("litellm_params", {}) or {}
-            # Resolve ``litellm_embedding_config`` here, at request-handling
-            # time, instead of at row-creation time. The resolved
-            # ``api_key`` / ``api_base`` / ``api_version`` lives only in
-            # this per-request ``data`` dict and is never persisted.
-            # Legacy rows that already carry a resolved (cleartext)
-            # ``litellm_embedding_config`` skip the lookup and pass through
-            # unchanged so the embed call keeps working.
-            embedding_model: Final = litellm_params.get("litellm_embedding_model")
-            if embedding_model and not litellm_params.get("litellm_embedding_config"):
-                from litellm.proxy.proxy_server import prisma_client
-
-                resolved_config: Final = await _resolve_embedding_config(
-                    embedding_model=embedding_model, prisma_client=prisma_client
-                )
-                if resolved_config:
-                    # Build a fresh dict via spread instead of mutating
-                    # ``litellm_params`` in place — the registry hands back
-                    # a reference to its cached object, so an in-place
-                    # update would persist the resolved cleartext into the
-                    # in-memory cache for the lifetime of the process.
-                    litellm_params = {
-                        **litellm_params,
-                        "litellm_embedding_config": resolved_config,
-                    }
-            data.update(litellm_params)
-    return data
+    if vector_store_to_run is None:
+        return data
+    if user_api_key_dict is not None:
+        await assert_user_can_access_vector_store(
+            vector_store=vector_store_to_run,
+            user_api_key_dict=user_api_key_dict,
+        )
+    return {**data, **(await build_request_data_from_managed_vector_store(vector_store_to_run))}
 
 
 @router.post(
@@ -130,6 +157,7 @@ async def vector_store_search(
     )
 
     data = await _read_request_body(request=request)
+    reject_caller_embedding_selection_params(payload=data, source="the search request body")
     data["vector_store_id"] = vector_store_id
 
     # Check for legacy vector store registry (non-managed vector stores)

litellm/proxy/vector_store_endpoints/management_endpoints.py

@@ -470,7 +470,7 @@ async def create_vector_store_in_db(
     # exposed every env-stored embedding-model credential on the
     # ``/vector_store/{new,info,update,list}`` responses. Keep the user's
     # raw ``litellm_embedding_model`` reference; resolution now happens in
-    # ``_update_request_data_with_litellm_managed_vector_store_registry``
+    # ``build_request_data_from_managed_vector_store``
     # at request-handling time so the cleartext config exists only in
     # per-request memory and never reaches the database.
     if litellm_params:
@@ -864,7 +864,7 @@ async def update_vector_store(
         # embedding-config auto-resolve previously persisted cleartext
         # credentials into the row; resolution now happens at request-
         # handling time in
-        # ``_update_request_data_with_litellm_managed_vector_store_registry``
+        # ``build_request_data_from_managed_vector_store``
         # so this row only ever stores the user-supplied
         # ``litellm_embedding_model`` reference.
         if "litellm_params" in update_data:

litellm/rag/main.py

@@ -14,6 +14,7 @@ import contextvars
 from collections.abc import Coroutine, Iterator
 from contextlib import contextmanager
 from functools import partial
+from types import MappingProxyType
 from typing import TYPE_CHECKING, Any, Final
 
 import httpx
@@ -50,6 +51,21 @@ INGESTION_REGISTRY: Final[dict[str, type[BaseRAGIngestion]]] = {
     "vertex_ai": VertexAIRAGIngestion,
 }
 
+# Only these retrieval_config keys are forwarded to vector_stores.asearch as
+# provider-specific params. The explicit allowlist keeps caller-controlled
+# connection overrides (api_base, api_key, ...) away from the search call,
+# where they could redirect store credentials to an attacker-chosen host.
+_FORWARDABLE_RETRIEVAL_CONFIG_KEYS: Final = frozenset(
+    {
+        "aws_region_name",
+        "vector_bucket_name",
+        "embedding_model",
+        "litellm_embedding_model",
+        "litellm_embedding_config",
+        "litellm_credential_name",
+    }
+)
+
 
 def get_ingestion_class(provider: str) -> type[BaseRAGIngestion]:
     """
@@ -224,13 +240,20 @@ async def _execute_query_pipeline(
         raise ValueError("No query found in messages for RAG query")
 
     # 2. Search vector store
+    # Forward allowlisted provider retrieval_config extras (region, embedding
+    # model, bucket, credential refs) to the search call; kwargs win on conflict.
+    provider_search_params: Final = MappingProxyType(
+        {k: v for k, v in retrieval_config.items() if k in _FORWARDABLE_RETRIEVAL_CONFIG_KEYS}
+    )
+    forwarded_search_params: Final = MappingProxyType({**provider_search_params, **kwargs})
     with _suppressed_sub_call_billing():
         search_response: Final = await litellm.vector_stores.asearch(
             vector_store_id=retrieval_config["vector_store_id"],
             query=query_text,
             max_num_results=retrieval_config.get("top_k", 10),
             custom_llm_provider=retrieval_config.get("custom_llm_provider", "openai"),
-            **kwargs,
+            router=router,
+            **forwarded_search_params,
         )
 
     search_provider: Final = retrieval_config.get("custom_llm_provider", "openai")

litellm/router.py

@@ -2360,7 +2360,7 @@ class Router:
     @overload
     async def acompletion(
         self, model: str, messages: list[AllMessageValues], stream: Literal[True, False] = False, **kwargs
-    ) -> CustomStreamWrapper | ModelResponse: 
+    ) -> CustomStreamWrapper | ModelResponse:
         ...
 
     # fmt: on
@@ -6410,8 +6410,6 @@ class Router:
             "responses",
             "generate_content",
             "generate_content_stream",
-            "vector_store_search",
-            "vector_store_create",
             "ocr",
             "search",
             "video_generation",
@@ -6435,6 +6433,8 @@ class Router:
             return sync_wrapper
 
         if call_type in (
+            "vector_store_search",
+            "vector_store_create",
             "vector_store_retrieve",
             "vector_store_list",
             "vector_store_update",
@@ -6446,11 +6446,16 @@ class Router:
                 client: object | None = None,
                 **kwargs,
             ):
-                if custom_llm_provider and "custom_llm_provider" not in kwargs:
-                    kwargs["custom_llm_provider"] = custom_llm_provider
-                if kwargs.get("model"):
-                    return self._generic_api_call_with_fallbacks(original_function=original_function, **kwargs)
-                return original_function(**kwargs)
+                provider_kwargs: Final = (
+                    MappingProxyType({**kwargs, "custom_llm_provider": custom_llm_provider})
+                    if custom_llm_provider and "custom_llm_provider" not in kwargs
+                    else MappingProxyType(kwargs)
+                )
+                if provider_kwargs.get("model"):
+                    return self._generic_api_call_with_fallbacks(original_function=original_function, **provider_kwargs)
+                if call_type == "vector_store_search":
+                    return original_function(**MappingProxyType({**provider_kwargs, "router": self}))
+                return original_function(**provider_kwargs)
 
             return vector_store_sync_wrapper
 
@@ -6626,6 +6631,7 @@ class Router:
                 return await self._init_vector_store_api_endpoints(
                     original_function=original_function,
                     custom_llm_provider=custom_llm_provider,
+                    call_type=call_type,
                     **kwargs,
                 )
             elif call_type in ("afile_delete", "afile_content"):
@@ -6666,6 +6672,7 @@ class Router:
         self,
         original_function: Callable,
         custom_llm_provider: str | None = None,
+        call_type: str | None = None,
         **kwargs,
     ):
         """
@@ -6684,6 +6691,13 @@ class Router:
                 **kwargs,
             )
 
+        # For search, pass the router so provider transforms can resolve
+        # router-managed embedding models (e.g. S3 Vectors query embeddings).
+        # The merge also overrides any client-supplied `router` key.
+        if call_type == "avector_store_search":
+            search_kwargs: Final = MappingProxyType({**kwargs, "router": self})
+            return await original_function(**search_kwargs)
+
         # Otherwise, call the original function directly
         return await original_function(**kwargs)
 

litellm/vector_stores/main.py

@@ -7,7 +7,7 @@ import builtins
 import contextvars
 from collections.abc import Coroutine, Mapping
 from functools import partial
-from typing import Final
+from typing import TYPE_CHECKING, Final
 
 import httpx
 
@@ -29,6 +29,9 @@ from litellm.types.vector_stores import (
 from litellm.utils import ProviderConfigManager, client
 from litellm.vector_stores.utils import VectorStoreRequestUtils
 
+if TYPE_CHECKING:
+    from litellm.router import Router
+
 ####### ENVIRONMENT VARIABLES ###################
 # Initialize any necessary instances or variables here
 base_llm_http_handler = BaseLLMHTTPHandler()
@@ -280,6 +283,7 @@ async def asearch(
     timeout: float | httpx.Timeout | None = None,
     # LiteLLM specific params,
     custom_llm_provider: str | None = None,
+    router: "Router | None" = None,
     **kwargs,
 ) -> VectorStoreSearchResponse:
     """
@@ -308,6 +312,7 @@ async def asearch(
             extra_body=extra_body,
             timeout=timeout,
             custom_llm_provider=custom_llm_provider,
+            router=router,
             **kwargs,
         )
 
@@ -347,6 +352,7 @@ def search(
     timeout: float | httpx.Timeout | None = None,
     # LiteLLM specific params,
     custom_llm_provider: str | None = None,
+    router: "Router | None" = None,
     **kwargs,
 ) -> VectorStoreSearchResponse | Coroutine[object, object, VectorStoreSearchResponse]:
     """
@@ -450,6 +456,7 @@ def search(
             timeout=timeout or request_timeout,
             _is_async=_is_async,
             client=kwargs.get("client"),
+            router=router,
         )
 
         return response

tests/test_litellm/llms/s3_vectors/vector_stores/test_s3_vectors_transformation.py

@@ -1,4 +1,4 @@
-from unittest.mock import MagicMock, Mock
+from unittest.mock import AsyncMock, MagicMock, Mock, patch
 
 import httpx
 import pytest
@@ -9,6 +9,18 @@ from litellm.llms.s3_vectors.vector_stores.transformation import (
 from litellm.types.vector_stores import VectorStoreSearchResponse
 
 
+def _mock_router(model_names, sync=False):
+    """Router mock serving the given embedding model names."""
+    router = MagicMock()
+    router.get_model_list.return_value = [{"model_name": name} for name in model_names]
+    embedding_response = Mock(data=[{"embedding": [0.1, 0.2, 0.3]}])
+    if sync:
+        router.embedding = MagicMock(return_value=embedding_response)
+    else:
+        router.aembedding = AsyncMock(return_value=embedding_response)
+    return router
+
+
 class TestS3VectorsVectorStoreConfig:
     def test_init(self):
         """Test that S3VectorsVectorStoreConfig initializes correctly"""
@@ -28,19 +40,174 @@ class TestS3VectorsVectorStoreConfig:
         url = config.get_complete_url(None, litellm_params)
         assert url == "https://s3vectors.us-west-2.api.aws"
 
-    def test_get_complete_url_missing_region(self):
-        """Test that missing region raises error"""
+    def test_get_complete_url_missing_region(self, monkeypatch):
+        """Missing region falls back to the default region (parity with ingestion)"""
+        monkeypatch.delenv("AWS_REGION_NAME", raising=False)
+        monkeypatch.delenv("AWS_REGION", raising=False)
         config = S3VectorsVectorStoreConfig()
-        litellm_params = {}
-        with pytest.raises(ValueError, match="aws_region_name is required"):
-            config.get_complete_url(None, litellm_params)
+        url = config.get_complete_url(None, {})
+        assert url == "https://s3vectors.us-west-2.api.aws"
+
+    def test_get_complete_url_uses_env_region(self, monkeypatch):
+        """Missing region param resolves from AWS_REGION_NAME env var"""
+        monkeypatch.setenv("AWS_REGION_NAME", "eu-west-1")
+        monkeypatch.delenv("AWS_REGION", raising=False)
+        config = S3VectorsVectorStoreConfig()
+        url = config.get_complete_url(None, {})
+        assert url == "https://s3vectors.eu-west-1.api.aws"
+
+    def test_get_complete_url_invalid_region_format(self):
+        """Invalid region format raises"""
+        config = S3VectorsVectorStoreConfig()
+        with pytest.raises(ValueError, match="Invalid AWS region format"):
+            config.get_complete_url(None, {"aws_region_name": "Bad_Region!"})
 
-    @pytest.mark.skip(reason="Requires embedding API call, tested in integration tests")
     def test_transform_search_request(self):
-        """Test search request transformation"""
-        # This test requires making an actual embedding API call
-        # It's better tested in integration tests
-        pass
+        """Full request-body transformation with a router-injected embedding"""
+        config = S3VectorsVectorStoreConfig()
+        mock_logging_obj = Mock()
+        mock_logging_obj.model_call_details = {}
+        router = _mock_router(["text-embedding-3-small"], sync=True)
+
+        url, request_body = config.transform_search_vector_store_request(
+            vector_store_id="test-bucket:test-index",
+            query="test query",
+            vector_store_search_optional_params={"max_num_results": 7},
+            api_base="https://s3vectors.us-west-2.api.aws",
+            litellm_logging_obj=mock_logging_obj,
+            litellm_params={},
+            extra_body=None,
+            router=router,
+        )
+
+        assert url == "https://s3vectors.us-west-2.api.aws/QueryVectors"
+        assert request_body == {
+            "vectorBucketName": "test-bucket",
+            "indexName": "test-index",
+            "queryVector": {"float32": [0.1, 0.2, 0.3]},
+            "topK": 7,
+            "returnDistance": True,
+            "returnMetadata": True,
+        }
+        assert mock_logging_obj.model_call_details["query"] == "test query"
+
+    @pytest.mark.asyncio
+    async def test_atransform_search_uses_router_for_virtual_model(self):
+        """Regression: router-served embedding models must resolve via the router,
+        not a bare litellm.aembedding call (which has no deployment credentials)."""
+        config = S3VectorsVectorStoreConfig()
+        mock_logging_obj = Mock()
+        mock_logging_obj.model_call_details = {}
+        router = _mock_router(["my-embedding-model"])
+
+        with patch("litellm.aembedding", new=AsyncMock()) as mock_bare_aembedding:  # test-quality-ok: guards that the bare-embedding path is not taken; dispatch seam is the behavior under test
+            url, request_body = await config.atransform_search_vector_store_request(
+                vector_store_id="test-bucket:test-index",
+                query="test query",
+                vector_store_search_optional_params={},
+                api_base="https://s3vectors.us-west-2.api.aws",
+                litellm_logging_obj=mock_logging_obj,
+                litellm_params={"embedding_model": "my-embedding-model"},
+                extra_body=None,
+                router=router,
+            )
+
+        router.aembedding.assert_awaited_once_with(model="my-embedding-model", input=["test query"])
+        mock_bare_aembedding.assert_not_awaited()
+        assert request_body["queryVector"]["float32"] == [0.1, 0.2, 0.3]
+        assert request_body["topK"] == 5  # default
+
+    @pytest.mark.asyncio
+    async def test_atransform_search_falls_back_when_router_does_not_serve_model(self):
+        """Router present but embedding_model is not a router deployment ->
+        bare litellm.aembedding keeps working (provider-prefixed + env creds stores)."""
+        config = S3VectorsVectorStoreConfig()
+        mock_logging_obj = Mock()
+        mock_logging_obj.model_call_details = {}
+        router = _mock_router(["some-other-model"])
+
+        mock_bare = AsyncMock(return_value=Mock(data=[{"embedding": [0.4, 0.5]}]))
+        with patch("litellm.aembedding", new=mock_bare):  # test-quality-ok: stubs the bare-embedding fallback whose request body the test asserts on
+            _, request_body = await config.atransform_search_vector_store_request(
+                vector_store_id="test-bucket:test-index",
+                query="test query",
+                vector_store_search_optional_params={},
+                api_base="https://s3vectors.us-west-2.api.aws",
+                litellm_logging_obj=mock_logging_obj,
+                litellm_params={"embedding_model": "azure/text-embedding-3-small"},
+                extra_body=None,
+                router=router,
+            )
+
+        mock_bare.assert_awaited_once_with(model="azure/text-embedding-3-small", input=["test query"])
+        router.aembedding.assert_not_awaited()
+        assert request_body["queryVector"]["float32"] == [0.4, 0.5]
+
+    @pytest.mark.asyncio
+    async def test_atransform_search_without_router_uses_bare_embedding(self):
+        """Backward compat: no router -> bare litellm.aembedding as before"""
+        config = S3VectorsVectorStoreConfig()
+        mock_logging_obj = Mock()
+        mock_logging_obj.model_call_details = {}
+
+        mock_bare = AsyncMock(return_value=Mock(data=[{"embedding": [0.6, 0.7]}]))
+        with patch("litellm.aembedding", new=mock_bare):  # test-quality-ok: stubs the bare-embedding fallback whose request body the test asserts on
+            _, request_body = await config.atransform_search_vector_store_request(
+                vector_store_id="test-bucket:test-index",
+                query="test query",
+                vector_store_search_optional_params={},
+                api_base="https://s3vectors.us-west-2.api.aws",
+                litellm_logging_obj=mock_logging_obj,
+                litellm_params={},
+                extra_body=None,
+            )
+
+        mock_bare.assert_awaited_once_with(model="text-embedding-3-small", input=["test query"])
+        assert request_body["queryVector"]["float32"] == [0.6, 0.7]
+
+    def test_transform_search_uses_router_for_virtual_model_sync(self):
+        """Sync twin: router-served embedding model resolves via router.embedding"""
+        config = S3VectorsVectorStoreConfig()
+        mock_logging_obj = Mock()
+        mock_logging_obj.model_call_details = {}
+        router = _mock_router(["my-embedding-model"], sync=True)
+
+        with patch("litellm.embedding", new=MagicMock()) as mock_bare_embedding:  # test-quality-ok: guards that the bare-embedding path is not taken; dispatch seam is the behavior under test
+            _, request_body = config.transform_search_vector_store_request(
+                vector_store_id="test-bucket:test-index",
+                query="test query",
+                vector_store_search_optional_params={},
+                api_base="https://s3vectors.us-west-2.api.aws",
+                litellm_logging_obj=mock_logging_obj,
+                litellm_params={"embedding_model": "my-embedding-model"},
+                extra_body=None,
+                router=router,
+            )
+
+        router.embedding.assert_called_once_with(model="my-embedding-model", input=["test query"])
+        mock_bare_embedding.assert_not_called()
+        assert request_body["queryVector"]["float32"] == [0.1, 0.2, 0.3]
+
+    def test_transform_search_without_router_uses_bare_embedding_sync(self):
+        """Sync twin: no router -> bare litellm.embedding as before"""
+        config = S3VectorsVectorStoreConfig()
+        mock_logging_obj = Mock()
+        mock_logging_obj.model_call_details = {}
+
+        mock_bare = MagicMock(return_value=Mock(data=[{"embedding": [0.8, 0.9]}]))
+        with patch("litellm.embedding", new=mock_bare):  # test-quality-ok: stubs the bare-embedding fallback whose request body the test asserts on
+            _, request_body = config.transform_search_vector_store_request(
+                vector_store_id="test-bucket:test-index",
+                query="test query",
+                vector_store_search_optional_params={},
+                api_base="https://s3vectors.us-west-2.api.aws",
+                litellm_logging_obj=mock_logging_obj,
+                litellm_params={},
+                extra_body=None,
+            )
+
+        mock_bare.assert_called_once_with(model="text-embedding-3-small", input=["test query"])
+        assert request_body["queryVector"]["float32"] == [0.8, 0.9]
 
     def test_transform_search_request_invalid_vector_store_id(self):
         """Test that invalid vector_store_id format raises error"""

tests/test_litellm/proxy/rag_endpoints/test_rag_endpoints.py

@@ -324,6 +324,127 @@ def test_rag_query_stream_returns_event_stream(client_internal_user):
     assert "data: [DONE]" in response.text
 
 
+def test_rag_query_merges_managed_store_params(client_internal_user):
+    """
+    Regression: /v1/rag/query must consult the managed vector store registry
+    (like the direct /v1/vector_stores/{id}/search endpoint does) so that
+    provider, region, embedding model, etc. don't have to be repeated in
+    retrieval_config. Pre-fix the registry was never read, so managed S3
+    Vectors stores failed with "aws_region_name is required".
+    """
+    import litellm
+    from litellm.types.utils import ModelResponse
+
+    mock_vector_store = {
+        "vector_store_id": "s3-store",
+        "custom_llm_provider": "s3_vectors",
+        "litellm_params": {
+            "aws_region_name": "eu-west-1",
+            "embedding_model": "my-embed",
+            "vector_bucket_name": "bkt",
+        },
+    }
+    mock_registry = MagicMock()
+    mock_registry.get_litellm_managed_vector_store_from_registry.return_value = mock_vector_store
+
+    mock_response = ModelResponse(
+        id="chatcmpl-test",
+        choices=[{"index": 0, "message": {"role": "assistant", "content": "hi"}, "finish_reason": "stop"}],
+        model="gpt-4o-mini",
+    )
+
+    with patch(  # test-quality-ok: aquery is the endpoint's downstream boundary; the forwarded config is what the test asserts
+        "litellm.proxy.rag_endpoints.endpoints.litellm.aquery",
+        new_callable=AsyncMock,
+        return_value=mock_response,
+    ) as mock_aquery, patch.object(litellm, "vector_store_registry", mock_registry), patch(  # test-quality-ok: seeds the managed-store registry the merge under test reads and grants access so real store resolution runs
+        "litellm.proxy.vector_store_endpoints.utils.can_user_access_vector_store",
+        new=AsyncMock(return_value=True),
+    ):
+        response = client_internal_user.post(
+            "/v1/rag/query",
+            json={
+                "model": "gpt-4o-mini",
+                "messages": [{"role": "user", "content": "hello"}],
+                "retrieval_config": {"vector_store_id": "s3-store"},
+            },
+        )
+
+    assert response.status_code == 200, response.json()
+    mock_aquery.assert_awaited_once()
+    forwarded_config = mock_aquery.await_args.kwargs["retrieval_config"]
+    assert forwarded_config["vector_store_id"] == "s3-store"
+    assert forwarded_config["custom_llm_provider"] == "s3_vectors"
+    assert forwarded_config["aws_region_name"] == "eu-west-1"
+    assert forwarded_config["embedding_model"] == "my-embed"
+    assert forwarded_config["vector_bucket_name"] == "bkt"
+
+
+def test_rag_query_store_params_win_over_user_retrieval_config(client_internal_user):
+    """Registry values must win over user-supplied retrieval_config keys so callers cannot override store credentials."""
+    import litellm
+    from litellm.types.utils import ModelResponse
+
+    mock_vector_store = {
+        "vector_store_id": "s3-store",
+        "custom_llm_provider": "s3_vectors",
+        "litellm_params": {"aws_region_name": "eu-west-1"},
+    }
+    mock_registry = MagicMock()
+    mock_registry.get_litellm_managed_vector_store_from_registry.return_value = mock_vector_store
+
+    mock_response = ModelResponse(
+        id="chatcmpl-test",
+        choices=[{"index": 0, "message": {"role": "assistant", "content": "hi"}, "finish_reason": "stop"}],
+        model="gpt-4o-mini",
+    )
+
+    with patch(  # test-quality-ok: aquery is the endpoint's downstream boundary; the forwarded config is what the test asserts
+        "litellm.proxy.rag_endpoints.endpoints.litellm.aquery",
+        new_callable=AsyncMock,
+        return_value=mock_response,
+    ) as mock_aquery, patch.object(litellm, "vector_store_registry", mock_registry), patch(  # test-quality-ok: seeds the managed-store registry the merge under test reads and grants access so real store resolution runs
+        "litellm.proxy.vector_store_endpoints.utils.can_user_access_vector_store",
+        new=AsyncMock(return_value=True),
+    ):
+        response = client_internal_user.post(
+            "/v1/rag/query",
+            json={
+                "model": "gpt-4o-mini",
+                "messages": [{"role": "user", "content": "hello"}],
+                "retrieval_config": {"vector_store_id": "s3-store", "aws_region_name": "us-east-1"},
+            },
+        )
+
+    assert response.status_code == 200, response.json()
+    forwarded_config = mock_aquery.await_args.kwargs["retrieval_config"]
+    assert forwarded_config["aws_region_name"] == "eu-west-1"
+
+
+@pytest.mark.parametrize(
+    "blocked_key",
+    ["embedding_model", "litellm_embedding_model", "litellm_embedding_config", "litellm_credential_name"],
+)
+def test_rag_query_rejects_caller_embedding_selection_params(client_internal_user, blocked_key):
+    """
+    Regression: a caller must not pick the embedding model or credential used at
+    search time. Those resolve through the Router with the proxy's credentials,
+    bypassing the key's model permissions, so they may only come from the
+    managed store's server-side registration.
+    """
+    response = client_internal_user.post(
+        "/v1/rag/query",
+        json={
+            "model": "gpt-4o-mini",
+            "messages": [{"role": "user", "content": "hello"}],
+            "retrieval_config": {"vector_store_id": "s3-store", blocked_key: "attacker-choice"},
+        },
+    )
+
+    assert response.status_code == 400, response.json()
+    assert blocked_key in str(response.json())
+
+
 EICAR = r"X5O!P%@AP[4\PZX54(P^)7CC)7}$EICAR-STANDARD-ANTIVIRUS-TEST-FILE!$H+H*"
 INGEST_REQUEST = '{"ingest_options":{"vector_store":{"custom_llm_provider":"openai"}}}'
 

tests/test_litellm/proxy/vector_store_endpoints/test_vector_store_endpoints.py

@@ -3158,3 +3158,35 @@ class TestAzureAIAnalyzeNamedIndexClassification:
             user_api_key_dict=self._team_member("analyze", ["read"]),
         )
         assert result is True
+
+
+@pytest.mark.parametrize(
+    "blocked_key",
+    ["embedding_model", "litellm_embedding_model", "litellm_embedding_config", "litellm_credential_name"],
+)
+def test_vector_store_search_rejects_caller_embedding_selection_params(blocked_key):
+    """
+    Regression: the search request body must not pick the embedding model or
+    credential used to embed the query. Those resolve through the Router with
+    the proxy's credentials, bypassing the key's model permissions, so they may
+    only come from the managed store's server-side registration.
+    """
+    from fastapi.testclient import TestClient
+
+    from litellm.proxy.auth.user_api_key_auth import user_api_key_auth
+    from litellm.proxy.proxy_server import app
+
+    mock_auth = UserAPIKeyAuth(user_id="test_internal_user", user_role=LitellmUserRoles.INTERNAL_USER.value)
+    original_overrides = app.dependency_overrides.copy()
+    app.dependency_overrides[user_api_key_auth] = lambda: mock_auth
+    try:
+        client = TestClient(app)
+        response = client.post(
+            "/v1/vector_stores/s3-store/search",
+            json={"query": "hello", blocked_key: "attacker-choice"},
+        )
+    finally:
+        app.dependency_overrides = original_overrides
+
+    assert response.status_code == 400, response.json()
+    assert blocked_key in str(response.json())

tests/test_litellm/rag/test_main.py

@@ -259,6 +259,135 @@ async def test_aquery_streaming_bills_sub_call_costs_into_final_event():
     assert standard_logging_object["response_cost"] >= 0.003
 
 
+@pytest.mark.asyncio
+async def test_aquery_forwards_provider_retrieval_config_and_router_to_search():
+    """
+    Regression: provider-specific retrieval_config keys (aws_region_name,
+    embedding_model, vector_bucket_name, ...) and the router must be forwarded
+    to the vector store search call. Pre-fix they were silently dropped, so
+    /v1/rag/query failed with provider config errors (e.g. S3 Vectors
+    "aws_region_name is required") even when the caller supplied them.
+    """
+    from unittest.mock import AsyncMock
+
+    from litellm.types.vector_stores import VectorStoreSearchResponse
+
+    router = litellm.Router(
+        model_list=[
+            {
+                "model_name": "gpt-4o-mini",
+                "litellm_params": {"model": "openai/gpt-4o-mini", "api_key": "test-key"},
+            }
+        ]
+    )
+
+    fake_search = AsyncMock(
+        return_value=VectorStoreSearchResponse(
+            object="vector_store.search_results.page", search_query="q", data=[]
+        )
+    )
+    with patch("litellm.vector_stores.asearch", new=fake_search):  # test-quality-ok: asearch is the boundary the forwarding contract under test targets
+        response = await litellm.aquery(
+            model="gpt-4o-mini",
+            messages=[{"role": "user", "content": "hello"}],
+            retrieval_config={
+                "vector_store_id": "bkt:idx",
+                "custom_llm_provider": "s3_vectors",
+                "top_k": 5,
+                "aws_region_name": "eu-west-1",
+                "embedding_model": "my-embed",
+                "vector_bucket_name": "bkt",
+            },
+            router=router,
+            mock_response="hi",
+        )
+
+    assert isinstance(response, ModelResponse)
+    fake_search.assert_awaited_once()
+    search_kwargs = fake_search.await_args.kwargs
+    assert search_kwargs["vector_store_id"] == "bkt:idx"
+    assert search_kwargs["custom_llm_provider"] == "s3_vectors"
+    assert search_kwargs["max_num_results"] == 5
+    assert search_kwargs["router"] is router
+    # provider-specific extras forwarded
+    assert search_kwargs["aws_region_name"] == "eu-west-1"
+    assert search_kwargs["embedding_model"] == "my-embed"
+    assert search_kwargs["vector_bucket_name"] == "bkt"
+    # consumed keys are not duplicated into the spread
+    assert "top_k" not in search_kwargs
+
+
+@pytest.mark.asyncio
+async def test_aquery_minimal_retrieval_config_forwards_no_extras():
+    """
+    A minimal retrieval_config must not leak consumed keys (or invent extras)
+    into the vector store search call.
+    """
+    from unittest.mock import AsyncMock
+
+    from litellm.types.vector_stores import VectorStoreSearchResponse
+
+    fake_search = AsyncMock(
+        return_value=VectorStoreSearchResponse(
+            object="vector_store.search_results.page", search_query="q", data=[]
+        )
+    )
+    with patch("litellm.vector_stores.asearch", new=fake_search):  # test-quality-ok: asearch is the boundary the forwarding contract under test targets
+        await litellm.aquery(
+            model="gpt-4o-mini",
+            messages=[{"role": "user", "content": "hello"}],
+            retrieval_config={"vector_store_id": "vs_test_123", "custom_llm_provider": "openai"},
+            mock_response="hi",
+        )
+
+    fake_search.assert_awaited_once()
+    search_kwargs = fake_search.await_args.kwargs
+    assert search_kwargs["vector_store_id"] == "vs_test_123"
+    assert search_kwargs["custom_llm_provider"] == "openai"
+    assert search_kwargs["router"] is None
+    leaked = {"top_k", "filters", "retrieval_filter", "aws_region_name", "embedding_model", "vector_bucket_name"}
+    assert not (leaked & set(search_kwargs.keys()))
+
+
+@pytest.mark.asyncio
+async def test_aquery_does_not_forward_connection_override_keys_to_search():
+    """
+    Only allowlisted retrieval_config keys may reach the vector store search
+    call. Caller-controlled connection overrides (api_base, api_key, arbitrary
+    extras) must be dropped, otherwise a caller could redirect store
+    credentials to an attacker-chosen host.
+    """
+    from unittest.mock import AsyncMock
+
+    from litellm.types.vector_stores import VectorStoreSearchResponse
+
+    fake_search = AsyncMock(
+        return_value=VectorStoreSearchResponse(
+            object="vector_store.search_results.page", search_query="q", data=[]
+        )
+    )
+    with patch("litellm.vector_stores.asearch", new=fake_search):  # test-quality-ok: asearch is the boundary the forwarding contract under test targets
+        await litellm.aquery(
+            model="gpt-4o-mini",
+            messages=[{"role": "user", "content": "hello"}],
+            retrieval_config={
+                "vector_store_id": "bkt:idx",
+                "custom_llm_provider": "s3_vectors",
+                "aws_region_name": "eu-west-1",
+                "api_base": "https://attacker.example.com",
+                "api_key": "attacker-key",
+                "arbitrary_extra": "nope",
+            },
+            mock_response="hi",
+        )
+
+    fake_search.assert_awaited_once()
+    search_kwargs = fake_search.await_args.kwargs
+    assert search_kwargs["aws_region_name"] == "eu-west-1"
+    blocked = {"api_base", "api_key", "arbitrary_extra"}
+    assert not (blocked & set(search_kwargs.keys()))
+
+
 def test_rag_call_types_are_registered():
     """
     query/aquery/ingest/aingest are @client-decorated entry points, so their

tests/test_litellm/test_router.py

@@ -7575,6 +7575,118 @@ async def test_acreate_batch_request_bedrock_tags_override_deployment_tags():
         assert mock_sign.call_args.kwargs["data"]["tags"] == request_tags
 
 
+@pytest.mark.asyncio
+async def test_avector_store_search_injects_router():
+    """
+    Regression: router.avector_store_search must pass the router down to the
+    SDK search call so provider transforms can resolve router-managed
+    embedding models (e.g. S3 Vectors query embeddings).
+    """
+    from litellm.types.vector_stores import VectorStoreSearchResponse
+
+    expected_response = VectorStoreSearchResponse(
+        object="vector_store.search_results.page", search_query="q", data=[]
+    )
+    mock_asearch = AsyncMock(return_value=expected_response)
+    # Router.__init__ binds asearch via a local import, so patch the module
+    # attribute before constructing the Router.
+    with patch("litellm.vector_stores.main.asearch", new=mock_asearch):  # test-quality-ok: the SDK call is the only place the injected router kwarg is observable
+        router = litellm.Router(
+            model_list=[
+                {
+                    "model_name": "gpt-3.5-turbo",
+                    "litellm_params": {"model": "openai/gpt-3.5-turbo", "api_key": "test-key"},
+                }
+            ]
+        )
+        search_response = await router.avector_store_search(
+            vector_store_id="v", query="q", custom_llm_provider="s3_vectors"
+        )
+
+    assert search_response is expected_response
+    mock_asearch.assert_awaited_once()
+    assert mock_asearch.await_args.kwargs["router"] is router
+
+
+@pytest.mark.asyncio
+async def test_avector_store_create_does_not_inject_router():
+    """The router injection is gated on the search call type: the create path
+    must keep calling the SDK without a router kwarg."""
+    expected_response = {"id": "vs_1", "object": "vector_store"}
+    mock_acreate = AsyncMock(return_value=expected_response)
+    # avector_store_create(model=None) resolves acreate via a local import at
+    # call time, so patching after Router construction works here.
+    router = litellm.Router(
+        model_list=[
+            {
+                "model_name": "gpt-3.5-turbo",
+                "litellm_params": {"model": "openai/gpt-3.5-turbo", "api_key": "test-key"},
+            }
+        ]
+    )
+    with patch("litellm.vector_stores.main.acreate", new=mock_acreate):  # test-quality-ok: the SDK call is the only place a leaked router kwarg would surface
+        create_response = await router.avector_store_create(model=None, custom_llm_provider="openai")
+
+    assert create_response is expected_response
+    mock_acreate.assert_awaited_once()
+    assert "router" not in mock_acreate.await_args.kwargs
+
+
+def test_vector_store_search_injects_router():
+    """
+    Sync parity for the router injection: router.vector_store_search must pass
+    the router down to the SDK search call so provider transforms can resolve
+    router-managed embedding models, same as avector_store_search.
+    """
+    from litellm.types.vector_stores import VectorStoreSearchResponse
+
+    expected_response = VectorStoreSearchResponse(
+        object="vector_store.search_results.page", search_query="q", data=[]
+    )
+    mock_search = MagicMock(return_value=expected_response)
+    # Router.__init__ binds search via a local import, so patch the module
+    # attribute before constructing the Router.
+    with patch("litellm.vector_stores.main.search", new=mock_search):  # test-quality-ok: the SDK call is the only place the injected router kwarg is observable
+        router = litellm.Router(
+            model_list=[
+                {
+                    "model_name": "gpt-3.5-turbo",
+                    "litellm_params": {"model": "openai/gpt-3.5-turbo", "api_key": "test-key"},
+                }
+            ]
+        )
+        search_response = router.vector_store_search(
+            vector_store_id="v", query="q", custom_llm_provider="s3_vectors"
+        )
+
+    assert search_response is expected_response
+    mock_search.assert_called_once()
+    assert mock_search.call_args.kwargs["router"] is router
+    assert mock_search.call_args.kwargs["custom_llm_provider"] == "s3_vectors"
+
+
+def test_vector_store_create_does_not_inject_router():
+    """The sync create path must keep calling the SDK without a router kwarg."""
+    expected_response = {"id": "vs_1", "object": "vector_store"}
+    mock_create = MagicMock(return_value=expected_response)
+    # Router.__init__ binds create via a local import, so patch the module
+    # attribute before constructing the Router.
+    with patch("litellm.vector_stores.main.create", new=mock_create):  # test-quality-ok: the SDK call is the only place a leaked router kwarg would surface
+        router = litellm.Router(
+            model_list=[
+                {
+                    "model_name": "gpt-3.5-turbo",
+                    "litellm_params": {"model": "openai/gpt-3.5-turbo", "api_key": "test-key"},
+                }
+            ]
+        )
+        create_response = router.vector_store_create(custom_llm_provider="openai")
+
+    assert create_response is expected_response
+    mock_create.assert_called_once()
+    assert "router" not in mock_create.call_args.kwargs
+
+
 class TestPreRoutingStrategyRegistryLifecycle:
     """
     Regression tests: a deployment leaving the model_list must release the

tests/test_litellm/vector_stores/test_main.py

@@ -0,0 +1,78 @@
+"""
+Tests for litellm/vector_stores/main.py.
+
+Pins the router threading contract for vector store search: the router is an
+explicit named parameter that reaches the HTTP handler, and it must never leak
+into litellm_params/kwargs where logging would model_dump() it (the #19550
+serialization trap).
+"""
+
+from unittest.mock import MagicMock, patch
+
+import litellm.vector_stores.main as vector_stores_main
+from litellm.vector_stores.main import search
+
+MOCK_SEARCH_RESPONSE = {
+    "object": "vector_store.search_results.page",
+    "search_query": "q",
+    "data": [],
+}
+
+
+def test_search_threads_router_to_handler():
+    """search() must pass its router param through to the HTTP handler"""
+    mock_router = MagicMock()
+    logger = MagicMock()
+
+    with (
+        patch(  # test-quality-ok: stubs provider config resolution; the seam under test is the router kwarg threading
+            "litellm.vector_stores.main.ProviderConfigManager.get_provider_vector_stores_config",
+            return_value=MagicMock(),
+        ),
+        patch.object(  # test-quality-ok: the handler call is the observable boundary for the router kwarg contract
+            vector_stores_main.base_llm_http_handler,
+            "vector_store_search_handler",
+            return_value=MOCK_SEARCH_RESPONSE,
+        ) as mock_handler,
+    ):
+        response = search(
+            vector_store_id="bkt:idx",
+            query="q",
+            custom_llm_provider="s3_vectors",
+            router=mock_router,
+            litellm_logging_obj=logger,
+        )
+
+    assert response == MOCK_SEARCH_RESPONSE
+    mock_handler.assert_called_once()
+    assert mock_handler.call_args.kwargs["router"] is mock_router
+
+
+def test_search_router_not_in_litellm_params():
+    """Regression (#19550 class): the router must stay out of GenericLiteLLMParams,
+    otherwise pre-call logging model_dump()s it and breaks serialization."""
+    mock_router = MagicMock()
+    logger = MagicMock()
+
+    with (
+        patch(  # test-quality-ok: stubs provider config resolution; the seam under test is litellm_params contents
+            "litellm.vector_stores.main.ProviderConfigManager.get_provider_vector_stores_config",
+            return_value=MagicMock(),
+        ),
+        patch.object(  # test-quality-ok: the handler call is where a leaked router in litellm_params would surface
+            vector_stores_main.base_llm_http_handler,
+            "vector_store_search_handler",
+            return_value=MOCK_SEARCH_RESPONSE,
+        ) as mock_handler,
+    ):
+        search(
+            vector_store_id="bkt:idx",
+            query="q",
+            custom_llm_provider="s3_vectors",
+            router=mock_router,
+            litellm_logging_obj=logger,
+        )
+
+    litellm_params = mock_handler.call_args.kwargs["litellm_params"]
+    assert "router" not in litellm_params.model_dump(exclude_none=True)
+    assert getattr(litellm_params, "router", None) is None

ui/litellm-dashboard/src/app/(dashboard)/vector-stores/_components/VectorStoreTester.test.tsx

@@ -116,16 +116,33 @@ describe("VectorStoreTester", () => {
     await waitFor(() => expect(mockSearch).toHaveBeenCalledTimes(1));
   });
 
-  it("reports a failed search and keeps the history empty", async () => {
+  it("shows the backend error in the history when a search fails", async () => {
     const user = userEvent.setup();
-    mockSearch.mockRejectedValue(new Error("boom"));
+    const errorBody = '{"error":{"message":"OpenAIException - api_key is required"}}';
+    mockSearch.mockRejectedValue(new Error(errorBody));
     renderTester();
 
     await user.type(queryInput(), "hello");
     await user.click(searchButton());
 
-    await waitFor(() => expect(mockFromBackend).toHaveBeenCalledWith("Failed to search vector store"));
-    expect(screen.getByText(EMPTY_STATE)).toBeInTheDocument();
+    await waitFor(() => expect(mockFromBackend).toHaveBeenCalledWith(errorBody));
+    expect(screen.getByText(`Search failed: ${errorBody}`)).toBeInTheDocument();
+    expect(screen.queryByText("No results found")).not.toBeInTheDocument();
+    expect(screen.queryByText(EMPTY_STATE)).not.toBeInTheDocument();
+    // the failed query stays in the input for retry
+    expect(queryInput()).toHaveValue("hello");
+  });
+
+  it('renders "No results found" for an empty result set, not an error', async () => {
+    const user = userEvent.setup();
+    mockSearch.mockResolvedValue({ object: "vector_store.search_results.page", search_query: "hello", data: [] });
+    renderTester();
+
+    await user.type(queryInput(), "hello");
+    await user.click(searchButton());
+
+    expect(await screen.findByText("No results found")).toBeInTheDocument();
+    expect(screen.queryByText(/search failed/i)).not.toBeInTheDocument();
   });
 
   it("clears the search history", async () => {

ui/litellm-dashboard/src/app/(dashboard)/vector-stores/_components/VectorStoreTester.tsx

@@ -40,6 +40,7 @@ export const VectorStoreTester: React.FC<VectorStoreTesterProps> = ({ vectorStor
     {
       query: string;
       response: VectorStoreSearchResponse | null;
+      error: string | null;
       timestamp: number;
     }[]
   >([]);
@@ -59,6 +60,7 @@ export const VectorStoreTester: React.FC<VectorStoreTesterProps> = ({ vectorStor
       const historyEntry = {
         query,
         response,
+        error: null,
         timestamp: Date.now(),
       };
 
@@ -66,7 +68,9 @@ export const VectorStoreTester: React.FC<VectorStoreTesterProps> = ({ vectorStor
       setQuery("");
     } catch (error) {
       console.error("Error searching vector store:", error);
-      toast.fromError("Failed to search vector store");
+      const errorMessage = error instanceof Error ? error.message : String(error);
+      toast.fromError(errorMessage);
+      setSearchHistory((prev) => [{ query, response: null, error: errorMessage, timestamp: Date.now() }, ...prev]);
     } finally {
       setIsLoading(false);
     }
@@ -228,7 +232,13 @@ export const VectorStoreTester: React.FC<VectorStoreTesterProps> = ({ vectorStor
                           })}
                         </div>
                       ) : (
-                        <div className="text-sm text-muted-foreground">No results found</div>
+                        <div
+                          className={
+                            entry.error ? "text-sm break-words text-destructive" : "text-sm text-muted-foreground"
+                          }
+                        >
+                          {entry.error ? `Search failed: ${entry.error}` : "No results found"}
+                        </div>
                       )}
                     </div>
                   </div>

ui/litellm-dashboard/src/components/networking.tsx

@@ -6970,7 +6970,7 @@ export const vectorStoreSearchCall = async (
     if (!response.ok) {
       const errorData = await response.text();
       await handleError(errorData);
-      return null;
+      throw new Error(errorData);
     }
 
     const data = await response.json();